search

skirdey / voicerestore

VoiceRestore: Flow-Matching Transformers for Universal Speech Restoration
MIT License
88 stars 9 forks source link

What data was the model trained on? #7

Closed DenisIvanko closed 1 month ago

DenisIvanko commented 1 month ago

Hello, dear @skirdey! Thanks for the nice work.

I didn't understand from the paper, what database of clean speech/noise was used to train the model? What data was the model trained on? Thank you!

skirdey commented 1 month ago

The data set was just clean speech sampled at 48 khz, noisy speech is being created on the fly from the same samples using VST plug-ins